[Mlir-commits] [mlir] [mlir][gpu] Add extra value types for gpu::ShuffleOp (PR #104605)

[Jakub Kuderski]

https://github.com/kuhar requested changes to this pull request.

Thanks for trying to make this more general. My main question is whether we want shuffle fully general and support any types of known bitwidth (other scalars, perhaps also vector types), or if we want to keep it close to what is supported in the hardware.
I think I'd be leaning more toward the former since the expansion to supported types is straightforward. We took a similar approach with `gpu.subgroup_reduce` that supports vector types. Might be wort aligning the two if we are at it, since `subgroup_reduce` can expand to shuffles. 

A few comments regarding the the PR description:

> gpu-to-vulkan: currently doesn't lower the shuffle op

I find this misleading, this is not a relevant conversion in this context.

> gpu-to-rocdl: only supports 32-bit anyway

This doesn't preclude us from expanding it 32-bit shuffles.


https://github.com/llvm/llvm-project/pull/104605